The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data in the form of code diffs, pull requests, and project files for review purposes.
Ingestion points: The workflow in SKILL.md involves reading and analyzing external code diffs and acceptance criteria provided by users or retrieved from version control.
Boundary markers: The instructions do not specify delimiters or explicit instructions to treat code comments or content as non-instructional data, which may leave the agent vulnerable to instructions embedded within a malicious diff.
Capability inventory: The skill mentions spawning and delegating tasks to "subagents" to achieve a "fresh-context" review.
Sanitization: There is no mention of sanitizing, escaping, or filtering the input code content before it is processed by the agent or its subagents.

referee